A comparison of object-based image analysis approaches for field boundary delineation using multi-temporal Sentinel-2 imagery

In this study we evaluated several Earth observation methodologies for automatically delineating crop field boundaries from multi-temporal Sentinel-2 imagery. The methodology makes use of edge detection that is applied to multiple images acquired during a growing season and image segmentation to delimit agricultural fields, orchards and vineyards. Two edge detection operators (Canny and Scharr) and three image segmentation algorithms (watershed, multi-threshold and multi-resolution) were combined and evaluated, resulting in six experiments (scenarios). A rule-based (knowledge-based) classification was applied to the extracted image objects to discard uncultivated areas. Reference field boundaries, manually digitised from very high (10 cm) resolution aerial imagery, were used for quantitative accuracy assessments and qualitative comparisons. The quantitative accuracy assessment consisted of both area- and edge-based metrics. The results showed that the watershed segmentation, combined with Canny edge detection, produced the most accurate field boundaries with an OA of 92.9% and combined MAE of 24.5 m. The Scharr algorithm produced thicker edges, causing positional errors along the boundaries. The multi-resolution and multi-threshold segmentation algorithms produced boundary inaccuracies ranging from 1 to 3 pixels, largely due to the creation of thick boundary objects, which caused offsets of the extracted borders on both sides of the reference boundaries. The combination of Canny edge detection (performed on multiple Sentinel-2 images acquired during the growing season) and watershed segmentation is thus recommended for operational field boundary delineation.